If you’re new, welcome! If you’re a regular, thanks for being here as we navigate this AI shift in eCommerce. I share what I’m testing and what’s actually moving the needle. Where’s the vault you ask? Every previous edition is saved here.

In this edition (shortcuts):
Your Skills Are Never Finished💥
Dots: OpenAI's Agent With Its Own Computer 🧰
Codex Has a Walk Away Button Too 📱
My Team Has Superpowers 🎤
News and Events Worth Following

The first skill in our official library went in on January 27.

Today there are 97 of them.

NONE of them are “finished”.

That is not a flaw. That is how a skill library is supposed to stay sharp.

A lot of people share freebie skills, it’s a cheap marketing trick to gain followers.

First of all, every skill needs customization to YOUR specific business need. So you have to invest in understanding how to build skills the right way.

The part nobody demos is the upkeep. A skill that was right in January can drift by March unless someone keeps tuning it.

Our team does the tuning. It is the real work behind how PPC Ninja runs advertising and delivers the best efficiency to our clients.

Since January, those 97 skills have changed 602 times.

That count leaves out six sweeping commits that touched ten or more skills at once, so a bulk cleanup is not padding it.

The middle skill has been changed 3 times.

Sixteen have been changed ten times or more.

A skill is not a document you write once. It is closer to a tool you keep sharpening. 72 of our 97 skills have scripts behind them.

Here are four from the last month, with the before and after.

Before: the skill described its method. "Statistical root cause analysis for Amazon Ads performance anomalies."

After: it opens with the moment you would reach for it. "Tells you which number actually broke on a bad day."

Nobody types "statistical root cause analysis" into a chat. They type "why did ACoS spike?" The label has to match the words people use.

Before: the skill listed what Amazon recommends for an account: products to advertise, bid suggestions, share of voice gaps.

After: it opens with the caveat. "Treat every one of these as a lead, never as an instruction."

Amazon's suggestions can ignore whether a product is in stock or where it sits in its life. Every item now gets checked before anyone acts on it.

Before: we renamed one skill folder. Four files still pointed at the old name.

After: a check scans every skill file when we commit, and fails the commit if a path points at a folder that does not exist.

The AI follows paths. A wrong one does not throw an error. It just sends the AI to the wrong place.

Before: our bulk file generator relied on the skill telling the AI how to set up each campaign.

After: the generator refuses to write a campaign if a required setting is missing, and it blocks aggressive launch settings unless a person explicitly asked for them.

Here is the one that taught me the most.

We had a skill that clearly said: before you build a bulk file, read the bulk file skill in full.

A second skill listed it as a related skill.

The AI shipped a flawed output.

Pointing it at the rule was not the same as following it.

This is when hallucinations happen.

So we stopped relying on just the instruction. Every generator now has to call a validator, and a check fails the commit if it does not.

That is the pattern. A rule written as advice gets skimmed. A rule written as a check gets enforced.

Beauty is one of the hardest shelves to defend because of relatively low customer loyalty and high openness to trying trendy TikTok brands or K-beauty alternatives. Dozens of shades and sizes. A rival can climb into your keyword fast.

So here is one of our skills at work that allows our team to pull competitive search rank on ANY tracked keywords (without even specifying who the competitors are). Getting this skill to where it stands took many turns of fine tuning and we are constantly updating it.

Our skills describe the goal in one line: easy to discover, cheap to load, hard to misuse. Each of those has a number behind it.

Claude does not read all 97 skills at the start of a chat.

It reads one short label from each, the part at the top of the skill file called front matter.

That is how it decides which skill to open.

All 97 labels together come to about 42,500 characters. That is roughly 10,000 tokens of Claude's memory used up before I type a word (my estimate: characters divided by four).

The median label is 392 characters. Two or three sentences.

So a label has one job: help Claude pick the right skill.

We write each one like a tab on a filing cabinet, not a summary of what is inside. It says when to reach for the skill, in the words you would actually say out loud.

Get the label wrong and a good skill never gets opened.

The body only loads when a skill matches.

Our bodies total about 163,000 words. The median is 1,342 words. The largest is nearly 11,000.

Every match pays for all of it.

You do not need 97 skills. You need one set of instructions you actually use.

Whether it is a Claude skill, a project or a long saved prompt, three habits:

🔶 Fix it the day it breaks. When the AI gets something wrong, decide whether the fix belongs in the instructions. If it does, write it that day.

🔶 Keep the label short. If a skill fires on the wrong job, the label is the problem.

🔶 Turn the must never rules into checks. If breaking a rule costs money, stop describing it and make a script verify it.

Back in August I wrote about making every mistake a rule. Rules need upkeep too.

How many sets of instructions are you running, and when did you last open them? Hit reply and tell me.

~Ritu

Dear reader: Do you need a well managed ad agency on your side? PPC Ninja has AI and tech in its DNA and is a pro at managing advertising and creatives for E-commerce. Hit reply if you or someone you know needs us.

Most of what we do with AI happens while we sit there.

You ask. It answers. You ask again.

On September 29, at DevDay, OpenAI announced something built for the opposite.

They are called Dots, and they come from OpenAI. They keep working after you walk away.

A Dot is OpenAI's personal agent, built into ChatGPT and powered by GPT-6 Astra.

Each one gets its own cloud computer and browser, so it keeps going when your own computer is closed.

You give it a goal, not a single prompt, and it works toward it in the background.

It connects to more than 4,000 apps through OpenAI's plugin ecosystem. You reach it from ChatGPT, Codex, Slack or Microsoft Teams. Here is everything OpenAI announced at DevDay in 15 minutes.

🔶 Its own computer. Nothing on your own computer has to stay on or awake.

🔶 Memory. A Dot builds memory from your conversations and the apps you connect.

🔶 Guardrails. Built in action rules, your own rules, automatic review of sensitive actions and activity monitoring.

🔶 Real work on stage. In the live demo, a Dot picked up a bug report from a channel, investigated it and opened a pull request. OpenAI says its engineers have Dots fixing dozens of bugs a day.

🔶 Disconnecting an app does not erase what the Dot already learned from it.

🔶 Reset erases everything. Conversations, memories and schedules go together.

🔶 The safety controls are new. Nobody has independently measured how often they fail yet.

🔶 Pricing is half known. Your first Dot is included with ChatGPT Pro and Business Premium, and talking to it does not eat into your usage. OpenAI has not published the allowance or the price for extra Dots.

🔶 Not everywhere yet. Pro access excludes the EEA, UK and Switzerland at launch. Texting a Dot is limited to a US Pro beta.

A Dot works with the identities, credentials and tools you assign it. Decide which logins it gets before you hand over any.

The Dots above run on OpenAI's computer.

This tip runs on yours.

My last issue covered the Claude side of this: You Don't Have to Watch It Work.

Codex Remote lets your phone drive a Codex session on your Mac. The Mac does the work. The phone is the remote.

On your Mac, open the ChatGPT app and go to:

Scan the QR code with your phone and sign in to the same ChatGPT account.

Both answer the same question. Neither is an autonomous agent running your business while you sleep.

Three switches decide whether it works once you leave:

🔶 Keep this Mac awake. Leave it off and the session dies when the Mac dozes. On a MacBook the lid has to stay open, or you need an external display.

🔶 Use your Mac apps while locked. Convenient, and the toggle to think hardest about.

🔶 Set up Chrome extension. It lets Codex navigate websites and fill out forms.

The work stays on your machine. Same files, same logins, same tabs. Anything your Mac can reach, your phone can now reach.

Two labs shipped the same feature. That is usually a signal.

I joined Michael Maher on The Longer Game to talk about what AI does to an agency team.

My team suddenly has superpowers. I'm NOT going to fire anyone from my team because of AI, like some of the other agencies are doing. We are an advertising and creative services agency for the AI era and we are proud to be heavy users of AI in our day-to-day operations.

I sat down with Hai Mag on Connected Commerce, the podcast from Eva Commerce. The idea: ecommerce teams do not need more SOPs. They need AI skills.

Nearly 4,000 selling partners met in Seattle. The source is Amazon's own recap.

🔶 Seller Assistant, upgraded. It remembers past decisions, watches your business, and recommends or acts within limits you set, with every action logged. You can describe tasks like restocking or pricing in plain language. Amazon says sellers act on its recommendations about 90 percent of the time. Free.

🔶 Seller Assistant for ads. It checks listing and inventory readiness, drafts Sponsored Products and Sponsored Brands campaigns, then keeps optimizing within the bid maximums you set. Amazon reports first sales 40 percent faster in early results. Available now.

🔶 Seller Central plugin for Claude and Amazon Quick. It carries your selling data and skills into the AI tool you already use. US first, in beta.

🔶 Alexa for Shopping. Amazon says more than 350 million customers used it last year, and shoppers who describe a problem in natural language spend over 40 percent more per order.

Listing Protection (US sellers by year end, worldwide in 2027), Ads Data Manager, selling on Walmart, eBay, TikTok Shop and Shopify from Seller Central, Amazon Business Deals, Amazon Supply Chain Services, and Amazon Lending open to every seller by year end.

Amazon Ads held unBoxed in San Francisco on September 30. The source is Amazon's announcement page.

🔶 Amazon Ads Agent. One conversational buying experience that brings the ads console and DSP together. Closed beta, expanding.

🔶 Full Funnel Campaigns. Set a goal and AI picks the channels, creative and audiences. In beta, Amazon reports 67 percent higher long term ROAS and 29 percent lower new customer acquisition cost.

🔶 DVA+. Display, streaming TV and audio in one campaign with one budget. Launched.

🔶 Generative AI creative. Display ads and TV quality video from your product assets, inside the Ads Agent. Select advertisers only.

🔶 Natural language reporting. Ask about roughly 1,000 metrics in plain language and get a business review in minutes.

🔶 Amazon Ads MCP Server (Lite). Starts agents with three core tools and reveals more as needed. Announced, still in development.

Also: AMC expert tools that write AMC SQL from a plain request, Branded Conversations inside Alexa for Shopping, and Sponsored Services with Yelp.

Every number above is Amazon's own claim from its announcements, not independent results.

Dots are in Cool Tools above. The rest of the keynote is below, with the time to jump to in the DevDay video.

🔶 ChatGPT Spaces (5:07). A shared page for you, your teammates and your Dots. Available today to Pro, Business Premium and Enterprise users.

🔶 Specialist Dots (8:04, preview). Virtual teammates a company sets up for work like accounting, marketing and legal. OpenAI is working with Microsoft to bring them into Agent 365.

🔶 GPT-6.1 Sol and Ultra Fast (8:38). Near Astra level intelligence at a fifth of the price, plus a mode that runs eight times faster at six times the standard price.

🔶 Codex in the cloud (10:44). Codex now runs fully in the cloud, and the Codex CLI got a refresh.

🔶 OpenAI Marketplace (14:57). Sell software to businesses using their existing OpenAI commitment. Over 30 launch partners.

The thread through all of it: agents that work on their own computer, with your access, while you do something else. At PPC Ninja we build on the same idea. AI does the heavy lifting under rules a team keeps tuning.

We hope you liked this edition of the AI for E-Commerce Newsletter! Hit reply and let us know what you think! Thank you for being a subscriber! Know anyone who might be interested to receive this newsletter? Share it with them and they will thank you for it! 😃 Ritu